Todd Deshane presented results from benchmarking tests of Xen and KVM virtualization systems at the 2008 Xen Summit. Key findings included Xen having similar or better CPU performance than KVM, but KVM outperforming Xen on some disk and network tests. Tests of performance isolation showed Xen was generally more isolated, while KVM scaling failed as guest numbers increased beyond 4. Areas for further work were identified, such as expanding tests and automating processes.

1. Todd Deshane, Ph.D. Student, Clarkson University
Xen Summit, June 23-24, 2008, Boston, MA, USA.

2. Xen and the Art of Virtualization (2003)
◦ Reported remarkable performance results
Xen and the Art of Repeated Research (2004)
◦ Validated performance results
Quantifying the Performance Isolation
Properties of Virtualization Systems (2007)
◦ Isolation Benchmark Suite
◦ Performance isolation testing methodology
◦ Lots of attention from virtualization developers and industry
Xen Summit Boston 2008

3. Understand architectural differences
◦ Stand-alone versus integrated hypervisor
Help developers realize areas of improvement
◦ Difficult for developers to test all cases
◦ Overall performance is important, but not the only factor
Help users make informed decisions
◦ Growing number of virtualization options to choose from
◦ Different users have different virtualization needs
◦ Hardware and software versions can make a big difference
Xen Summit Boston 2008

4. Base machine
◦ Ubuntu Linux 8.04 AMD64
Software packages from Ubuntu repositories
◦ Linux kernel 2.6.24-18
◦ Xen 3.2.1+2.6.24-18-xen
◦ KVM 62
Guests
◦ Ubuntu Linux 8.04 AMD64
◦ Automated debootstrap
Hardware
◦ 2.4 GHz Intel Core 2 CPU 6600, 4 GB of RAM, 250 GB disk
Xen Summit Boston 2008

5. Overall performance
◦ Performance of the whole system and components
◦ Focused on macro-benchmarks
◦ Standard benchmarks and repeatable methods
Performance isolation
◦ Protection from resource consumption from other guests
◦ Representative workload with and without stress tests
Scalability
◦ Ability to run more guests without loss of performance
◦ Same workload on each guest, increase number of guests
Xen Summit Boston 2008

6. Macro-benchmarks
◦ CPU
◦ Kernel compile
◦ Disk I/O
Automated guest build
◦ Benchvm virtualization benchmark suite
Automated testing and reporting
◦ Phoronix Test Suite
Xen Summit Boston 2008

7. Xen Summit Boston 2008

8. Xen and KVM had similar CPU performance
◦ Xen: 0.999, KVM: 0.993
Xen was better than KVM on kernel compile
◦ Xen: 0.487, KVM: 0.384
KVM was better on disk I/O
◦ Write – Xen: 0.855, KVM: 0.934
◦ Read – Xen: 0.852, KVM: 0.994
◦ Disk caching effects?
Xen Summit Boston 2008

9. Isolation Benchmark Suite
◦ Memory stress test: calloc()
◦ Fork stress test: fork()
◦ CPU stress test
◦ Mixed calculations in tight loop
◦ Disk
◦ Threaded IOzone read and write
◦ Network receiver
◦ Receive threaded UDP traffic from external host
◦ Network sender
◦ Send threaded UDP traffic to external host
Xen Summit Boston 2008

10. Xen Summit Boston 2008

11. Xen Summit Boston 2008

12. Xen was isolated on memory, fork, CPU, disk
Xen was slightly isolated on network sender
Xen showed no isolation on network receiver
◦ Kernel bug(s) in Ubuntu?
Xen had unexpectedly good disk performance
KVM was well-isolated on all stress tests
KVM had unexpectedly good network sender
performance
KVM had unexpectedly poor disk and network
receiver performance
Xen Summit Boston 2008

13. Ran Apache compile in 1, 2, 4, 16, and 30
guests
Measured compile time and number of guests
that ran to completion
Xen Summit Boston 2008

14. Xen Summit Boston 2008

15. Xen scaled linearly with respect to number of
guests
KVM had many guest crashes
4 guests: 1 crashed guest
◦
8 guests: 4 crashed guests
◦
16 guests: 7 crashed guests
◦
30 guests: system crashed during compile
◦
Xen Summit Boston 2008

16. Virtualization benchmarking is still difficult
Testing on multiple categories is crucial
Automated testing is important and useful
Transparency in testing methods
◦
Repeatability is needed, yet challenging
◦
Always more cases to test
◦
Challenging to adequately benchmark rapidly evolving
◦
technologies
Xen Summit Boston 2008

17. Extend testing to include Xen HVM, and KVM
with paravirt I/O
More complete automation of testing process
with benchvm
Port benchvm to Python
Add support for more distros in benchvm
Use Phoronix Test Suite-like functionality
Test profiles, test suites, batch benchmarking
◦
Automated results parsing
◦
Graphing/uploading of results
◦
Automated system and test config collection and publishing
◦
Xen Summit Boston 2008

18. Benchmarking co-researchers
◦ Zachary Shepherd, Jeanna Neefe Matthews, Muli Ben-
Yehuda, Amit Shah, and Balaji Rao
Performance isolation and scalability
researchers
◦ Wenjin Hu and Madhujith Hapuarachchi
Early developers and testers of benchvm
◦ Cyrus Katrak and Martin McDermott
Members of the Xen and KVM communities
◦ Feedback and support
Xen Summit Boston 2008

19. Xen Summit Boston 2008